Topic modeling analysis of conversations across various stages of dementia

With the growing use of artificial intelligence (AI)-powered conversational technology, people with dementia may benefit from these tools. Finetuning and adapting such systems require an understanding of the language use of people with dementia and communication patterns.

Daily conversations with 120 Korean older adults at various stages of dementia were collected over a two-year period and transcribed. Text mining techniques were applied to identify language patterns and latent meanings, including term frequency (TF), term frequency-inverse document frequency (TF-IDF), co-occurrence network and concordance analyses, and topic modeling using both latent Dirichlet allocation (LDA) and BERT-based Topic Modeling (BERTopic).

TF/TF-IDF analyses of 6,989 speech segments by people with dementia revealed frequent use of pronouns and vague demonstratives. Word pairs and in-context usage of frequently used words were closely tied to their lived experiences. Topic modeling identified five themes via LDA and nine key topics from 29 BERTopic subcategories, covering hardship, family and relationships, emotions, reminiscence, and daily pleasures. BERTopic also captured health-related needs, identity, and agency, offering insight into the inner world of people with dementia.

Despite declining cognitive abilities, people with dementia experience a wide range of life events and emotions. This study highlights the need for language-based support to help them retain agency, rather than be viewed solely as dependent or in need of care. Furthermore, the words and themes from this study could serve as valuable input for developing AI-powered communication algorithms, enabling more dementia-friendly and engaging conversations with people with dementia.